Masked Topology Modeling for Self-Supervised Learning on Parametric CAD
TLDR
Self-supervised pretraining on parametric CAD using masked topology modeling on B-Rep face-adjacency graphs, combined with contrastive learning, achieves strong benchmark performance.
Reasoning
Strengths include a novel self-supervised task that leverages the unique B-Rep structure, combining masked edge prediction with contrastive learning and augmentation. Weaknesses are that the abstract lacks specific downstream task details and explicit comparisons to prior methods, though benchmarks are mentioned.
Read-first score
Read-first score 58.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 72.
Field roles
Rank sensitivity
Stability: volatile; rank range: 8.
Keyword Scores
Deep Analysis
Innovations
- Masked Topology Modeling (MTM): a self-supervised pretraining task that masks edges in the face-adjacency graph of B-Reps and predicts each masked edge's convexity and curve type.
- Combination of MTM with MoCo-style momentum-queue contrastive learning and B-rep-aware augmentations.
- A BFS-connected face-region masked-reconstruction objective.
- A new procedurally generated dataset for pretraining (alongside ABC dataset).
Methodology
The method pretrains a graph encoder on the face-adjacency graph of CAD boundary representations (B-Reps). MTM masks edges and uses a prediction head to recover convexity and curve type from face features. It is combined with a momentum-queue contrastive objective using B-rep-specific augmentations and a masked region reconstruction objective. Pretraining uses the ABC dataset and a new procedurally generated dataset, with evaluation on multiple downstream benchmarks.
Key Results
The proposed pretraining approach shows strong performance on a number of benchmarks.